DOI: 10.1126/science.1238303, (2013);341 Science

et al.Sandro SantagataSupports the Anabolic Malignant StateTight Coordination of Protein Translation and HSF1 Activation

This copy is for your personal, non-commercial use only.

clicking here.colleagues, clients, or customers by , you can order high-quality copies for yourIf you wish to distribute this article to others

  here.following the guidelines

can be obtained byPermission to republish or repurpose articles or portions of articles

  ): November 24, 2014 www.sciencemag.org (this information is current as of

The following resources related to this article are available online at

http://www.sciencemag.org/content/341/6143/1238303.full.htmlversion of this article at:

including high-resolution figures, can be found in the onlineUpdated information and services,

http://www.sciencemag.org/content/suppl/2013/07/17/341.6143.1238303.DC1.html can be found at: Supporting Online Material

http://www.sciencemag.org/content/341/6143/1238303.full.html#relatedfound at:

can berelated to this article A list of selected additional articles on the Science Web sites

http://www.sciencemag.org/content/341/6143/1238303.full.html#ref-list-1, 17 of which can be accessed free:cites 46 articlesThis article

http://www.sciencemag.org/content/341/6143/1238303.full.html#related-urls10 articles hosted by HighWire Press; see:cited by This article has been

http://www.sciencemag.org/cgi/collection/cell_biolCell Biology

subject collections:This article appears in the following

registered trademark of AAAS. is aScience2013 by the American Association for the Advancement of Science; all rights reserved. The title

CopyrightAmerican Association for the Advancement of Science, 1200 New York Avenue NW, Washington, DC 20005. (print ISSN 0036-8075; online ISSN 1095-9203) is published weekly, except the last week in December, by theScience

on

Nov

embe

r 24

, 201

4w

ww

.sci

ence

mag

.org

Dow

nloa

ded

from

o

n N

ovem

ber

24, 2

014

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fr

om

on

Nov

embe

r 24

, 201

4w

ww

.sci

ence

mag

.org

Dow

nloa

ded

from

o

n N

ovem

ber

24, 2

014

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fr

om

on

Nov

embe

r 24

, 201

4w

ww

.sci

ence

mag

.org

Dow

nloa

ded

from

o

n N

ovem

ber

24, 2

014

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fr

om

on

Nov

embe

r 24

, 201

4w

ww

.sci

ence

mag

.org

Dow

nloa

ded

from

o

n N

ovem

ber

24, 2

014

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fr

om

on

Nov

embe

r 24

, 201

4w

ww

.sci

ence

mag

.org

Dow

nloa

ded

from

o

n N

ovem

ber

24, 2

014

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fr

om

on

Nov

embe

r 24

, 201

4w

ww

.sci

ence

mag

.org

Dow

nloa

ded

from

o

n N

ovem

ber

24, 2

014

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fr

om

Tight Coordination of Protein Translation and HSF1 Activation Supports the Anabolic Malignant StateSandro Santagata, Marc L. Mendillo, Yun-chi Tang, Aravind Subramanian, Casey C. Perley,

Stéphane P. Roche, Bang Wong, Rajiv Narayan, Hyoungtae Kwon, Martina Koeva,

Angelika Amon, Todd R. Golub, John A. Porco Jr., Luke Whitesell,* Susan Lindquist*

Introduction: Ribosome biogenesis is commonly up-regulated to satisfy the increased anabolic demands associated with malignant transformation and tumor growth. Many different oncogenic sig-naling pathways converge on the ribosome to increase translational fl ux. Despite the detailed under-standing of ribosome regulation in cancer, it is not clear whether the net translational activity of the ribosome can itself regulate transcriptional programs that support and promote the malignant state.

Methods: To investigate the transcriptional effects of modulating translational activity in malignant cells, we used integrated chemical and genetic approaches, including a gene signature–based genetic and chemical screen of more than 600,000 gene expression profi les (LINCS database) and an indepen-dent, reporter-based chemical screen of more than 300,000 compounds. A lead compound was tested in several cell lines unifi ed by their increased dependence on HSF1 activation for growth and survival, and in an in vivo cancer model.

Results: Inhibiting translation led to large changes in the transcriptome. The single most enriched category consisted of genes regulated by the heat-shock transcription factor, HSF1. The most down-regulated mRNA was HSPA8, which encodes the constitutive HSP70 chaperone that helps to fold nascent polypeptides. The expression of many other genes that HSF1 coordinates to support cancer was also strongly affected. HSF1 protein levels were unchanged, but HSF1 DNA occupancy was nearly eliminated. Inhibition of the HSF1-regulated gene expression program is thus a dominant transcrip-tional effect elicited by inhibiting protein translation.

Using a gene signature of HSF1 inactivation to query the LINCS database revealed a strong con-nection between HSF1 inactivation and perturbations that inhibit protein translation, including a broad spectrum of chemical and genetic interventions that target the ribosome, eukaryotic initiation factors (eIFs), aminoacyl tRNA synthetases, and upstream signaling/regulatory pathways that control translation.

Our high-throughput small-molecule screen identifi ed rocaglamide A, an inhibitor of translation initiation, as the strongest inhibitor of HSF1 activation. An analog of this compound, RHT, increased thioredoxin-interacting protein (TXNIP) mRNA and protein levels and decreased glucose uptake and lactate production. Cell-based cancer models characterized by high dependence on HSF1 activation for growth and survival were highly sensitive to RHT, as were cells derived from diverse hematopoietic malignancies. RHT had a strong antitumor effect—with marked inhibition of HSF1 activity and glucose uptake—against xenografted acute myeloid leukemia cells.

Discussion: The ribosome functions as a central information hub in malig-nant cells: Translational fl ux conveys information about the cell’s metabolic status to regulate the transcriptional programs that support it. Multiple unbi-ased chemical and genetic approaches establish HSF1 as a prime transducer of this information, centrally poised to regulate the transcription of genes that support protein folding, biomass expansion, anabolic metabolism, cellular pro-liferation, and survival. Targeting translation initiation may offer a strategy for reversing HSF1 activation, disabling metabolic and cytoprotective pathways in malignant cells.

FIGURES IN THE FULL ARTICLE

Fig. 1. Inhibiting protein translation

inactivates HSF1.

Fig. 2. LINCS analysis reveals that targeting

protein translation inactivates HSF1.

Fig. 3. Chemical screens reveal that targeting

translation control inactivates HSF1.

Fig. 4. Inhibiting translation initiation with

rocaglates ablates HSF1 DNA binding.

Fig. 5. Rocaglates modulate tumor energy

metabolism.

Fig. 6. Rocaglates selectively target aneuploid

cancer cells and nontransformed cells with

cancer-associated genetic aberrations.

Fig. 7. Rocaglates suppress tumor growth,

HSPA8 mRNA levels, and glucose uptake

in vivo.

SUPPLEMENTARY MATERIALS

Materials and MethodsFigs. S1 to S9Tables S1 to S5References and Notes

RELATED ITEMS IN SCIENCE

V. Gandin, I. Topisirovic, Trans-HFS1 express. Science 341, 242-243 (2013). DOI: 10.1126/science.1242359

HSF1 at the crossroads of protein translation and metabolism. (Left) Cancers activate an HSF1-regulated transcriptional program to adapt to the anabolic demands of relentless biomass expansion. Glucose uptake increases, and expression of TXNIP, an inhibitor of glucose uptake, drops. (Right) Down-regulating translation with rocaglate scaffold initiation inhibitors reverses cancer-associated HSF1 activation. Glucose uptake drops as TXNIP levels rise.

19 JULY 2013 VOL 341 SCIENCE www.sciencemag.org 250

RESEARCH ARTICLE SUMMARY

READ THE FULL ARTICLE ONLINE

http://dx.doi.org/10.1126/science.1238303

Cite this article as S. Santagata et al., Science 341, 1238303 (2013). DOI: 10.1126/science.1238303

The list of author affi liations is available in the full article online.*Corresponding author. E-mail: whitesell@wi.mit.edu (L.W.); lindquist_admin@wi.mit.edu (S.L.)

Published by AAAS

Tight Coordination of ProteinTranslation and HSF1 ActivationSupports the Anabolic Malignant StateSandro Santagata,1,2,3* Marc L. Mendillo,3,4* Yun-chi Tang,4,5† Aravind Subramanian,6

Casey C. Perley,3,4 Stéphane P. Roche,7 Bang Wong,6 Rajiv Narayan,6 Hyoungtae Kwon,3,4

Martina Koeva,3,4 Angelika Amon,4,5 Todd R. Golub,6 John A. Porco Jr.,7

Luke Whitesell,3‡ Susan Lindquist3,4‡

The ribosome is centrally situated to sense metabolic states, but whether its activity, in turn,coherently rewires transcriptional responses is unknown. Here, through integrated chemical-geneticanalyses, we found that a dominant transcriptional effect of blocking protein translation incancer cells was inactivation of heat shock factor 1 (HSF1), a multifaceted transcriptional regulatorof the heat-shock response and many other cellular processes essential for anabolic metabolism,cellular proliferation, and tumorigenesis. These analyses linked translational flux to the regulationof HSF1 transcriptional activity and to the modulation of energy metabolism. Targeting this linkwith translation initiation inhibitors such as rocaglates deprived cancer cells of their energy andchaperone armamentarium and selectively impaired the proliferation of both malignant andpremalignant cells with early-stage oncogenic lesions.

Regulation of ribosome activity is crit-ical for supporting cellular prolifera-tion. In cancer, ribosome biogenesis is

commonly increased to satisfy the increasedanabolic demands associated with malignanttransformation and tumor growth (1–4). In ad-dition, many different oncogenic signaling path-ways converge on the ribosome to modulateits function (5, 6). These inputs are integrated,and the net ribosomal translational activity istuned to reflect the metabolic and proliferativestate of the cell. Despite the detailed under-standing of ribosome regulation in cancer, it isnot well understood whether the net trans-lational activity of the ribosome can itself beconveyed to regulate transcriptional programsthat support and promote the malignant state.Is a coherent and coordinated transcriptionalresponse triggered by modulating translationactivity?

Results

Inhibiting Translational Flux Inactivates HSF1To investigate the transcriptional effects of re-ducing translational flux through the ribosome inmalignant cells, we analyzed the mRNA expres-sion profiles of breast cancer cells after treatmentwith various inhibitors of translation elongation(anisomycin, emetine, cephaeline, and cyclohex-imide). Large changes in the transcriptome werehighly correlated across all four inhibitors [Pearsoncorrelation coefficient (r) between 0.85 to 0.97for all pairwise correlations]. The most stronglyenriched category consisted of genes regulatedby promoters that contain DNA binding motifsfor the heat-shock transcription factor known asheat shock factor 1 (HSF1) (P = 9.87 × 10–7)(Fig. 1A and table S1). Of the 13,258 genes mea-sured, the single most down-regulated mRNAwasHSPA8, which encodes a constitutive HSP70chaperone that folds nascent polypeptides as theyemerge from the ribosome (Fig. 1B and table S2)(7).HSPA1A, a cancer-inducedHSP70 gene, wasalso among the 10 most down-regulated mRNAs.This transcriptional response suggested that re-duced translational flux causes a profound shift inthe activity of HSF1.

In a wide range of cancers, HSF1 regulates atranscriptional network that is distinct from theconventional network activated by thermal stress(8). This cancer network includes many classic“heat-shock” genes. But, it also includes a broadcadre of other genes that play critical roles in ma-lignancy, some of which are positively regulatedby HSF1 and some negatively regulated. All fourinhibitors of translation elongation affected genesin the HSF1 cancer network (P = 0.016) (Fig. 1Cand fig. S1). Genes that are positively regulated

by HSF1 were down-regulated when translation-al flux through the ribosome was reduced. Thesegenes included drivers of cell proliferation andmitogenic signaling (such asCENPA,CKS1B, andPRKCA), transcription and mRNA processing(such as LSM2 and LSM4), protein synthesis(such as FXR1 andMRPL18), energymetabolism(such asMAT2A, SLC5A3, PGK1,MBOAT7, andSPR), and invasion/metastasis (such asEMP2 andLTBP1). In a complementary fashion, genes thatare negatively regulated byHSF1were up-regulatedwhen translational flux through the ribosomewasreduced. These included genes that promote dif-ferentiation (such asNOTCH2NL), cellular adhe-sion (such asEFEMP1 andLAMA5), and apoptosis(such as BCL10, CFLAR, and SPTAN1).

This effect of translation inhibition on HSF1-regulated transcription led us to examine thegenome-wide pattern of DNA occupancy byHSF1in breast cancer cells. After a 6-hour exposure tocycloheximide, we performed chromatin immu-noprecipitation coupled with massively parallelDNA sequencing (ChIP-Seq) using a previouslyvalidated antibody against HSF1 (8). Despite cy-cloheximide treatment, HSF1 protein levels them-selves remained unchanged (Fig. 1D). In contrastto DNAoccupancy byRNA-polymerase II (whichwas not globally reduced), HSF1 occupancy wasnearly eliminated (Fig. 1, E toG, fig. S2, and tableS3). This held true for genes that were eitherpositively or negatively regulated byHSF1, aswellas for genes shared with the classic heat-shockresponse and genes specific to the HSF1 cancerprogram (table S3). Together, these data pointedto a link between the activity of the ribosome andthe activity of HSF1.

LINCS Establishes Translation as a PotentRegulator of HSF1 in Cancer CellsTo further investigate the link between the HSF1activity and translational program, we turned to anexpression-profiling resource created by the Li-brary of Integrated Network-based Cellular Sig-natures (LINCS) program (Fig. 2 and supplementarymaterials, materials and methods). The LINCSdatabase is a large catalog of gene-expressionprofiles collected from human cells treated withchemical (small-molecule) and genetic [short hair-pin RNA (shRNA)] perturbations.

We generated a query signature for HSF1 in-activation from expression profiles of breast can-cer cells that had been treatedwithHSF1 shRNAs(supplementary materials, materials and methods)(8). This signature included both genes that wereup-regulated by HSF1 inactivation and down-regulated by HSF1 inactivation.We compared ourHSF1 query signature with LINCS expressionprofiles from nine cell lines that are currently themost extensively characterized in this database(Fig. 2A). Eight of these are cancer lines of di-verse histopathologic origin. These lines have beentreated individuallywith 3866 small-molecule com-pounds or 16,665 shRNAs targeting 4219 genes.The compounds used for these gene expression

RESEARCHARTICLE

1Department of Pathology, Brigham and Women’s Hospital(BWH), and Harvard Medical School, Boston, MA 02215,USA. 2Dana Farber Cancer Institute, Boston, MA 02215, USA.3Whitehead Institute for Biomedical Research (WIBR), Cam-bridge, MA 02142, USA. 4Howard Hughes Medical Institute,Department of Biology, Massachusetts Institute of Technology(MIT), Cambridge, MA 02142, USA. 5David H. Koch Institutefor Integrative Cancer Research and Howard Hughes MedicalInstitute, MIT, Cambridge, MA 02142, USA. 6Broad Institute ofMIT and Harvard, Cambridge, MA 02142, USA. 7Departmentof Chemistry, Center for Chemical Methodology and LibraryDevelopment (CMLD-BU), Boston University, Boston, MA02215, USA.

*These authors contributed equally to this work.†Present address: Institute of Health Sciences, Shanghai Insti-tutes for Biological Sciences, Chinese Academy of Sciences &Shanghai Jiao Tong University School of Medicine, 20025Shanghai, China.‡Corresponding author. E-mail: whitesell@wi.mit.edu (L.W.);lindquist_admin@wi.mit.edu (S.L.)

www.sciencemag.org SCIENCE VOL 341 19 JULY 2013 1238303-1

profiles encompassed U.S. Food and Drug Ad-ministration (FDA)–approved drugs and knownbioactives. The shRNAs usedwere directed againstthe known targets of these compounds, againstgenes in related pathways, or against other genesthat have been implicated in a variety of humandiseases. In all, we compared our HSF1 signa-ture with 161,636 LINCS signatures, each gen-erated from at least three replicates (for a total of614,216 profiles.)

As expected, the LINCS perturbations thatnegatively correlated with our HSF1 inactivationsignature were enriched for known activators ofHSF1. They included shRNAs that target com-ponents of the proteasome. They also includedcompounds that inhibit the proteasome and thatinhibit HSP90 (Fig. 2, B and C and table S4).

The LINCS perturbations that positively cor-related with our HSF1 inactivation signature weremost highly enriched for translation inhibitors(cephaeline, cycloheximide, and emetine) (Fig. 2,B and C, and table S4). These perturbations werealso highly enriched for compounds that target

signaling pathways that regulate protein transla-tion: phosphatidylinositol 3-kinase (PI3Kinase)/mammalian target of rapamycin (mTOR) inhib-itors (Fig. 2B and table S4). Of the nearly 200gene ontology classes analyzed, the ribosome sub-unit family was the single most enriched (Fig. 2,B and C, and table S4). In addition, eukaryoticinitiation factors (eIFs) and aminoacyl tRNAsynthetases were also highly enriched. This un-biased analysis using the LINCS database dem-onstrates the connection between translational fluxand the function of HSF1 in cancer.

An Unbiased High-Throughput ChemicalScreen for HSF1 InhibitorsTo find alternate ways to inhibit HSF1, we per-formed a large high-throughput chemical screen.We screened 301,024 compounds through theU.S. National Institutes of Health (NIH) Molec-ular Libraries Probe Center Network (MLPCN;Pubchem AID, 2118) (Fig. 3A) using an HSF1-regulated reporter driven by consensus heat-shockelements (HSEs). To accommodate constraints of

the high-throughput 384-well format (supplemen-tary materials, materials and methods), we useda reporter cell line stably transduced with aluminescence-based reporter and induced HSF1activation with a simple proteotoxic stressor (theproteasome inhibitor MG132).

Approximately 2500 hit compounds from theprimary screen, which blocked induction of thereporter, were then counter-screened with an in-dependent dual reporter cell line (Fig. 3B) so asto eliminate nonselective inhibitors. This secondline had been stably transducedwith two constructs,one encoding a green fluorescent protein (GFP)driven by HSEs and the other encoding a red flu-orescent protein (RFP) driven by a doxycycline-regulated control promoter. Compounds thatselectively inhibit HSF1 activity should suppressGFP expression in this cell line but should notsuppress doxycycline-mediated induction of RFP.Notably, compounds that are thought to selec-tively inhibit HSF1—such as triptolide, quercetin,KNK423, and KNK437 (9)—all suppressedboth reporters (fig. S3). Thus, these compounds

Fig. 1. Inhibiting protein transla-tion inactivates HSF1. (A) Gene setenrichment analysis was performedby using the MSigDB web service re-lease version 3.84 (www.broadinstitute.org/gsea/index.jsp) on genes nega-tively regulated in breast cancer cellsfollowing a 6-hour incubation with in-hibitors of protein translation elongation. Selected results are displayed;complete GSEA results are provided in table S1. (B) Scatter plot of levels ofmRNA transcripts (log2) after a 6-hour incubation with the indicated inhibitorsof protein translation elongation versus control dimethyl sulfoxide (DMSO). Thelevels of HSPA1A and HSPA8 are indicated for each elongation inhibitor. (C)Translation elongation inhibitors alter the basal transcriptional program ofHSF1 in breast cancer cells. Genes bound by HSF1 in MCF7 were ranked bytheir differential expression between cells treated with translation elongationinhibitors (TI) and control DMSO (D). Each column represents a gene and isnormalized across the column, with high expression in red and low expressionin blue. (D) An immunoblot shows the levels of HSF1 protein and the loading

control tubulin after a 6-hour exposure to the indicated concentrations ofcycloheximide (CHX). (E) Heat map of RNA polymerase II ChIP-Seq read densityin MCF7 cells that were treated with DMSO or 10 mM CHX for 6 hours. Genomicregions from –2 kb to +2 kb relative to the transcription start site for all RefSeqgenes are shown. (F) Heatmap of HSF1 ChIP-Seq read density inMCF7 cells thatwere treated with DMSO or 10 mM CHX for 6 hours. Genomic regions from–1 kb to +1 kb relative to the peak of HSF1 binding for all HSF1-enrichedregions (union of all HSF1-enriched regions in the four data sets depictedhere) are shown. (G) Representative genes bound by HSF1 inMCF7 cells (HSPA8,HSPA1A, CKS2, and RBM23). The x axis depicts from −2kb from the transcriptionstart site (TSS) to 5 kb from the TSS for each gene.

19 JULY 2013 VOL 341 SCIENCE www.sciencemag.org1238303-2

RESEARCH ARTICLE

are far less specific for HSF1 than commonlyassumed.

More to the point, the chemical screen resultsalso suggested a link between HSF1 activation andthe translation machinery. By far the most potentand selective hit to emerge from the 301,024compounds we tested was the rocaglate knownas rocaglamide A [median inhibitory concentra-tion (IC50) of ~50 nM for the heat shock reporterversus IC50 >1000 nM for the control reporter](Fig. 3C). This natural product inhibits the func-tion of the translation initiation factor eIF4A, aDEAD box RNA helicase (10, 11). Presumably,it passed counterscreening in our secondary assaywith the dual reporter system because translationof the doxycycline-regulated RFP control doesnot require the classical cap-dependent initiationcomplex.

To define structure-activity relationships forinhibition of the HSE reporter by rocaglamide A,we used our dual reporter system to test 38

additional rocaglates (fig. S4). These includedboth natural products and totally synthetic ana-logs prepared with photocycloaddition methods(12, 13). Five hydroxamate analogs were morepotent than rocaglamide A at inhibiting the HSEreporter while retaining similar selectivity (tableS5). The most potent inhibitor had an IC50 of~20 nM (table S5). We named this compound[previously reported as “8e” (13)] Rohinitib (orRHT), for Rocaglate Heat Shock, Initiation ofTranslation Inhibitor.

Characterizing the Effects of RHT onCancer CellsTo validate findings from our engineered reportersystem, we measured the effects of RHT on thebasal expression of several endogenous HSF1-regulated transcripts (Fig. 3D and figs. S5 andS6). RHT did not reduce the transcript levels ofthe control housekeeping genesB2M andGAPDH.Nor did it reduce the transcript levels of HSF1

itself (Fig. 3D and fig. S6A). However, mRNAlevels of Hsp40 (DNAJA1) and Hsp70 genes(HSPA1B andHSPA8) dropped substantially. Themost dramatically affected was the constitutivelyexpressed HSPA8 gene (>90% reduction) (Fig.3D). This was also the gene that we had foundto be the most strongly repressed by translationelongation inhibitors (Fig. 1B).

The effects of RHTwere not due to reductionsin HSF1 protein levels, which remained constant(Fig. 3E and fig. S6B). The sharp decrease inHSP70 mRNA levels in response to RHT heldtrue across a histologically diverse panel of hu-man cancer cell lines (MCF7, breast adenocarci-noma;MO91,myeloid leukemia;CHP100, sarcoma;and HeLa, cervical carcinoma) as well as in ar-tificially transformed 293T kidney cells (Fig. 3Dand fig. S6, A and C). RHT had a much smallereffect on HSP70 mRNA levels in proliferatingbut nontumorigenic diploid cells (WI38 and IMR90)(fig. S6C).

+

-

-0.70 -0.50 -0.30 -0.10 0.10 0.30 0.50 0.70

Allperturbagens

Ribosomesubunits

Translationinhibitors

Proteasomeinhibitors

HSP90inhibitors

161k

442

142

56

390

Connectivity score

TotalsignaturesC Positive connectionsNegative connections

Null

+

0

-

1 161k 1 161k

Negativeconnection

Positiveconnection

HSF1 query signature of up/down regulated genes

Results ordered by connectivity to HSF1

Analytics calculateconnectivity score

161K unique gene expression signatures

BA NES Score

-3.05Proteasome inhibitors *HSP90 inhibitors *

EIF subunits *

Proteasome subunits *

2.28

-2.80

2.00

-1.20

1.71

1.261.35

PI3K/mTOR inhibitors *Ribosome subunits *

Aminoacyl tRNA synthetases *

Translation inhibitors *

0

Genes ranked by change in expression vs. control

* p < 10-2

Fig. 2. LINCS analysis reveals that targeting protein translation in-activates HSF1. (A) Schematic representation of the LINCS analysis usedto identify chemical and genetic modulators that are correlated with HSF1inactivation (supplementary materials, materials and methods). Pink repre-sents genes whose levels increase, and green represents genes whose levelsdecrease, after shRNA-mediated knockdown of HSF1. (B) GSEA results of ourHSF1 inactivation signature LINCS analysis. Perturbation signatures were rank-ordered by connectivity with the HSF1 inactivation signature, and enrichmentwas determined for KEGG pathway gene sets and ATC chemical classes (de-tails are available in the supplementary materials, materials and methods).

Normalized enrichment score (NES) of selected results are plotted (completeGSEA results are provided in table S4). (C) Barcode plot of the connectivityscore of all of the individual perturbations comprising the indicated enrichedchemical or gene sets. The bagel plot in the center of the barcode plotsummarizes the positive, negative, and null (not connected) fractions for theindicated enriched class. All perturbations that are positively or negativelyconnected for the indicated enriched classes are shown. Total perturbations ineach class are indicated on the right of the plot. Blue represents negativelyconnected, and red represents positively connected, classes of enrichedperturbations.

www.sciencemag.org SCIENCE VOL 341 19 JULY 2013 1238303-3

RESEARCH ARTICLE

To obtain a more direct and global view ofRHT’s effects on HSF1 activity, we examinedgenome-wide promoter occupancy by means ofChIP-Seq analysis. RHT virtually abolishedHSF1binding throughout the genome (Fig. 4, A and B;fig. S6D; and table S3). As had occurred withcycloheximide, RHTaffected both genes that arepositively regulated by HSF1 and genes that arenegatively regulated by HSF1 (Fig. 4A). Further-more, it affected both classic heat-shock genesand genes specific to the HSF1 cancer program(table S3). The effects on HSF1 DNA occupancyoccurred at concentrations of cycloheximide andRHT that inhibit the ribosome activity to a sim-ilar extent (Fig. 4C).

Rocaglates Modulate Tumor Energy MetabolismWhile characterizing the effects of RHT on thetranscriptome, we noted that treated cells failed toacidify the culture medium (detected incidentallyby the color of the pH indicator phenol red in-cluded in standard media). This suggested a re-versal of the “Warburg effect,” a metabolic shiftresponsible for increased lactic acid productionby many cancers. Genetic compromise of HSF1drives a shift in metabolism in both cell cultureand animal models (14, 15). Hence, this effect ofRHT is consistent with inactivation of HSF1.

Our mRNA expression profiling of rocaglate-treated breast cancer cells also revealed thatmRNA levels for thioredoxin-interacting protein

(TXNIP) were up-regulated. TXNIP is a power-ful negative regulator of glucose uptake and is awell-established regulator of cellular energy status(16, 17). Its expression is dramatically reducedin malignant cells, leading to increased glucoseuptake (18). Conversely, increasing TXNIP lev-els leads to reduced glucose uptake (16). The in-duction of TXNIP mRNA by RHTwas observedacross a diverse panel of tumor cell lines (Fig. 5A).TXNIP protein levels also increased sharply de-spite a marked reduction in the levels of othershort-lived proteins, such as p53 (Fig. 5B). Al-though we did not detect HSF1 bound to theTXNIP locus, HSF1 null cells showed higherlevels of TXNIP (fig. S7). In addition, HSF1 diddirectly regulate a group of other genes involvedin energymetabolism (includingMAT2A, SLC5A3,and PGK1). At a functional level, the effects ofRHTwere associatedwith concentration-dependentreductions in both glucose uptake and lactateproduction (Fig. 5C). Thus, the effects of RHTonprotein translation, HSF1 activation, and energymetabolism—processes lying at the core of theanabolic state of cancer—appear to be coordinated.

Rocaglates Selectively Target Pre-MalignantCells with Early-Stage Oncogenic LesionsDoes this tight coordination create vulnerabilitiesfor the malignant phenotype that could be ex-ploited as a therapeutic strategy? We looked at arange of cell-based cancer models unified by their

increased dependence on HSF1 activation forgrowth and survival. Although it occurs veryearly during oncogenesis, simple loss of the tu-mor suppressor Nf1 leads to an increase in HSF1protein levels, nuclear localization, and transcrip-tional activation (19).We treatedmouse embryonicfibroblasts (MEFs) in which Nf1 was knockedout and wild-type littermate control MEFs inwhich HSF1 was not activated, with either RHTor with cycloheximide. The two cell types weresimilarly sensitive to cycloheximide. However,Nf1-null MEFs were more sensitive than werewild-type MEFs to RHT (Fig. 6A). In this modelfor an early event in tumorigenesis, targeting trans-lation initiation rather than translation elongationseems to provide a more selective, better toleratedapproach for disrupting the link between trans-lation and HSF1 activation.

A second engineered system allowed us to askwhether rocaglates would selectively inhibit thegrowth of cells carrying a simple chromosomalaberration that models another common early eventin the development of cancer: aneuploidy. Chro-mosomal imbalances lead to both increased ener-gy and proteotoxic stress. This is reflected by theelevation of the HSF1-regulated chaperone pro-teinHSP72, encoded byHSPA1A (20).We isolatedMEFs from mice carrying Robertsonian fusionsfor chromosome 13 (21). TheseMEFs (TS13) car-ry a single extra copy of 120 Mb of chromosome13, introducing an additional copy of 843 genes.

Fig. 3. Chemical screens reveal that targetingtranslation control inactivates HSF1. (A) Flow-chart outlining the steps in the high-throughputMLPCN screen for inhibitors of HSF1 activation. (B)

Schematic of dual-reporter cell line used to counter-screen primary screen hits. GFP expression is regulated by a heat shock–inducible promoter. RFP expressionis regulated by a doxycycline response element (TetR). (C) Effect of rocaglamide A on the HSE-driven GFP and doxycycline-driven RFP signals after incubationwith 2.5 mM MG132 and 2 mg/ml doxycycline. Chemical structure of rocaglamide A is displayed in the inset. Error bars indicate mean T SEM of quadruplicates.(D) Effect of RHT on HSF1-regulated and control endogenous mRNA transcript levels in M0-91 leukemia cells measured by means of nanostring nCounter after a6-hour incubation with indicated concentrations of RHT. Levels of endogenous transcript are shown as percent of DMSO-treated control. Mean of duplicates arereported. (E) HSF1 protein levels are not affected in M0-91 leukemia cells treated with RHT. Immunoblot shows the levels of HSF1 protein and the loading control(Tubulin) after a 6-hour exposure to the indicated concentrations of RHT.

19 JULY 2013 VOL 341 SCIENCE www.sciencemag.org1238303-4

RESEARCH ARTICLE

Cycloheximide, as well as conventional cy-totoxic chemotherapeutics (taxol and hydroxyurea),inhibited the growth of both trisomic and litter-mate control MEFs to an equal extent (Fig. 6Band fig. S8). But, trisomicMEFs (P< 0.0001) weremore sensitive than wild-type MEFs to RHT(Fig. 6B). Thus, again in this model for an earlyneoplastic change that activates HSF1, targetingtranslation initiation seems to provide a bettertolerated, more selective approach for targetingthe malignant state.

HSF1 activation is even more prominent inadvanced malignancies (8, 22, 23). For example,colon cancers frequently show immunohistochem-ical evidence of strong HSF1 activation (Fig. 6C)and this correlates with poor clinical outcome (8).We mined publicly available expression profilingfrom colon cancer lines with highly aneuploidkaryotypes [chromosomal instability (CIN)] andfrom colon cancer lines with near-euploid kar-yotypes but microsatellite instability (MIN). TheCIN lines expressed markedly higher levels ofHSPA1A, which is consistent with greater levelsof proteotoxic stress and greater activation of theHSF1-regulated cancer program (Fig. 6, D andE). Next, we tested several patient-derived coloncancer lines with CIN and several patient-derivedcolon cancer lines with MIN for sensitivity toinhibition by RHT. The CIN lines were muchmore sensitive than were the MIN lines. Non-

transformed colon epithelial cell lines with eu-ploid chromosome content were the least sensitiveof all the lines we tested (Fig. 6F).

Rocaglates Suppress the Growth of CancerCells and TumorsSome rocaglates have previously been shown toexert profound anticancer activity (10, 24–26).We tested RHT against a collection of cell linesincluding nontransformed diploid lines andcancer cell lines with diverse histopathologicalorigins and oncogenic lesions (Fig. 7A). The non-transformed cell lines were relatively resistant toRHT (IC50 from 100 to 300 nM). All cancer celllines were sensitive to RHT (IC50 < 30 nM); thehematopoietic tumor cell lines were especiallysensitive (IC50 ≤ 6 nM). We used one of thesehematopoietic tumor lines, the M0-91 cell lineoriginally derived from a patient with acute mye-loid leukemia (27), to further characterize theeffects of RHT. RHTstrongly suppressedHSPA8mRNA levels inM0-91 cells and induced TXNIPmRNA (Fig. 7B). In addition, RHT sharply de-creased glucose uptake by these cells (Fig. 7C).

Are the effects of RHT in cell culture achiev-able at drug exposures that are systemically tol-erable in animals? To directly address this criticalissue of therapeutic index, we first used standardin vitro assays to test whether RHT had suffi-ciently drug-like properties to justify testing in

mice (fig. S9). We assessed aqueous solubility,plasma stability, plasma protein binding, hepaticmicrosome stability, and cellular permeability(fig. S9A). No severe liabilities were found. Wenext established minimally toxic parameters fordosing mice with RHT and performed a plasmapharmacokinetic study after administration of1 mg/kg subcutaneously (fig. S9, B and C). Peakplasma levels were far in excess of those requiredfor the key biological activities we had demon-strated in cell culture. Moreover, levels requiredfor anticancer activity in vitro were maintained inexcess of 2 hours in vivo.

We next established subcutaneous tumor xeno-grafts of the human myeloid leukemia cell lineM091 in nonobese diabetic (NOD)–severe com-bined immunodeficient (SCID) immunocompro-misedmice.When themean tumor volume reached100 mm3, we administered RHTat 1 mg/kg for 4consecutive days eachweek for 3 weeks (Fig. 7D).Over the treatment period, there was no evidenceof gross systemic toxicity (fig. S9D). RHTmediatedmarked, sustained inhibition of the growth of thisvery aggressive myeloid malignancy (Fig. 7D).

We then pursued pharmacodynamics studies.Mice bearing xenografts were given a single doseof RHT. Tumors were explanted 4 hours later,andHSPA8 and TXNIPmRNA levels were deter-mined by means of reverse transcription polymer-ase chain reaction (RT-PCR) (Fig. 7E). Similar to

Fig.4. Inhibitingtranslationinitiation with rocaglatesablatesHSF1DNAbinding.(A) Heat map of HSF1 ChIP-Seq read density inM0-91 cellsthat were treated with DMSO,20 nM RHT, 100 nM RHT, or10 mM CHX for 6 hours. Ge-nomic regions from –1 kb to+1 kb relative to the peak ofHSF1 binding for all HSF1 en-riched regions (union of allHSF1-enriched regions in theseven data sets depicted here)are shown. (B) RepresentativeHSF1-bound genes in M0-91cells (HSPA8, HSPA1B, CKS2,

and RBM23). The x axis depicts from –2 kb from the transcription start site (TSS) to 5 kb from the TSS for each gene. (C) Autoradiograph of S35-labeled proteinlysates from MCF7 cells treated for 6 hours with the indicated concentrations of RHT or CHX. Graphs show the counts per minute from acetone precipitation ofproteins in each sample, quantitated by using a scintillation counter.

www.sciencemag.org SCIENCE VOL 341 19 JULY 2013 1238303-5

RESEARCH ARTICLE

the effects we observed in cell culture, RHTcauseda strong decrease in HSPA8 transcript levels and astrong increase in TXNIP transcript levels. In aseparate experiment, we monitored the uptake offluorescently labeled 2-deoxyglucose 48 hoursafterRHTdosing. RHTstrongly suppressed uptakeof this glucose analog by these tumors (Fig.7F). Clearly, the effects of RHT achieved in cellculture could also be achieved in whole animals.

DiscussionHSF1 provides essential support for the malig-nant state by blocking apoptotic responses andpromoting protein synthesis, anabolic energymetabolism, mitogenic signaling pathways, andpathways that facilitate invasion and metastasis(8, 14, 15, 19, 23, 28–30). Here, we show that theability of HSF1 to maintain this cancer programis exquisitely sensitive to translational activity.

Our work establishes that the ribosome couldfunction as a central information hub: Transla-tional flux conveys information about the cell’smetabolic status to regulate the transcriptionalprograms that support it. The specific molecularmechanisms are likely to bemultifaceted, butHSF1is clearly a linchpin in this process. One plausiblemechanism for the effects of translation inhibitorson HSF1 activity could involve the translation ofmRNA for a modifier of HSF1 transcriptionalactivity that is sensitive to subtle changes in trans-lation activity, involving eIF4A and/or other initia-tion factors. BecauseHSF1 regulates the expressionof genes encoding for ribosomal subunits andother regulators of translation (8, 15, 31), the ex-istence of a feedforward regulatory circuit involv-ing the protein translation machinery and HSF1is also possible.

HSF1 is centrally poised to support proteinfolding and biomass expansion as well as manyother functions to which malignant cells areaddicted (8, 14, 15, 32). We postulate that thetranslation/HSF1 link we have uncovered incancer may derive from ancient systems gearedto align and synchronize essential cellular func-tions for growth and survival. In comparison, in thenematode HSF1 is a longevity factor, and in yeastit is an essential gene that participates in cotransla-tional quality control (33–35).

In man, the translation/HSF1 link is particu-larly important in supporting the malignant phe-notype because it can respond to varied metabolicinputs that are commonly dysregulated in cancer(5, 6, 36–38). This translation/HSF1 link allowsthese metabolic inputs to bolster the cytoprotec-tive milieu, helping tumor cells to accommodatethe drastic internal imbalances arising during on-cogenesis as well as the severe external stressesarising from therapeutic interventions (39). Thetight coordination of protein translation and HSF1activation, together with the many ways that cellsintegrate the derangements of malignancy withtranslational activity, suggests that unifying prin-ciples drive HSF1 activation across the extraor-dinarily wide range of human cancers in whichthat activation occurs (8, 22).

Although cancer cells often co-opt powerful,adaptive nononcogene systems for their benefit(40), it now appears that by co-opting the link be-tween the ribosome and HSF1, cancers becomeespecially vulnerable to agents that target transla-tion and its upstream regulatory pathways. In thisregard, our animal experiments suggest that tar-geting translation initiation may offer a strategyfor reversing HSF1 activation and disabling themetabolic and cytoprotective addictions of ma-lignant cells.

Materials and Methods

Cell LinesWI38, CHP100, HeLa, 293T, PC3, MCF7, andNIH3T3 cells were purchased from AmericanType Culture Collection (ATCC). ImmortalizedNf1 knockout MEFs and littermate wild-type con-trol MEFs were kind gifts from K. Cichowski.Littermate-derived euploid and trisomic primaryMEFs were described previously (20). RHT treat-ments experiments were performed by using chro-mosome 13 trisomic cell lines and littermatecontrol euploid cell lines that carried a singleRobertsonian translocation. Early-passage MEFs

were used to ensure that additional karyotypicchanges had not yet occurred. Two primary hu-man cell lines (CCD112CoN andCCD841CoN),five MIN lines (HCT-116, HCT-15, DLD-1,SW48, and LoVo), and five CIN lines (Caco2,HT-29, SW403, SW480, and SW620) were ob-tained from ATCC. Chromosome number andkaryotype information were obtained from the Na-tional Cancer Institute database and the COSMICdata set at the Sanger Institute. M0-91 cells werepreviously described (27). The M0-91 cell lineused in this studywere established from explantedM0-91 tumors that had been xenografted once inmice. All cell cultures were maintained under 5%CO2 in media according to their specifications.

mRNA Expression Profiling and AnalysisExpression profiles for MCF7 cells treated for6 hourswith anisomycin (15 mM), emetine (7 mM),cephaeline (6 mM), and cycloheximide (14 mM)were previously deposited in the ConnectivityMap (41). MCF7 cells were treated with 200 nMrocaglamide A or 50 nM RHT for 6 hours, andRNA was then purified after extraction withTRIzol reagent [Invitrogen (Carlsbad, CA), cata-log no. 15596-026]. Gene expression analysis was

Fig. 5. Rocaglatesmodulatetumor energy metabolism.(A) TXNIPmRNA transcript lev-els in a panel of cancer celllines measured by nanostringnCounter after a 6-hour incu-bation with 50 nM RHT. Meanof duplicates are reported. (B)Immunoblot showing TXNIPlevels in the indicated cancercell lines after a 6-hour incu-bation with the indicated con-

centration of RHT. b-actin is the loading control. The effect on p53, a short half-life protein, is shown. (C)Effects of the indicated amount of RHT on [H3]-2-deoxyglucose uptake (left) and lactate production (right)in a panel of cancer cell lines after 8 hours of incubation. Error bars indicate mean T SEM of triplicates.

19 JULY 2013 VOL 341 SCIENCE www.sciencemag.org1238303-6

RESEARCH ARTICLE

performed by using Affymetrix (Santa Clara, CA)GeneChip HTHumanGenomeU133A 96-ArrayPlates, and data was analyzed as previously de-scribed (8). All microarray raw data were de-

posited in a public database [National Center forBiotechnology Information Gene ExpressionOmnibus (GEO) pending]. Gene set enrichmentanalysis of the differentially expressed genes

after treatment of MCF7 cells with translationelongation inhibitors was performed by usingthe gene set enrichment analysis (GSEA) Website release version 3.84 (42). Enrichment for

Fig. 6. Rocaglates selectively target aneuploid cancer cells and nontransformedcells with cancer-associated genetic aberrations. (A) Photomicrographs ofNf1wild typeand Nf1 null MEFs that were treated for 14 days with 25 nM RHT. The relative viable cellnumber of RHT-treated (middle) and CHX-treated (right) are shown. Error bars indicatemean T SEM of n = 6 replicates. (B) Effect of 72 hours treatment with either RHT (left) orcycloheximide (right) on the proliferation of MEFs carrying a single extra copy of 120 Mb of chromosome 13 (TS13) compared with MEFs derived from littermatecontrols (WT), [mean T SD, n = 3 replicates, ***P < 0.001, two-way analysis of variance (ANOVA)]. (C) Photomicrographs of normal colon epithelial cells andinvasive colon adenocarcinoma (hematoxylin and eosin stains and HSF1 immunohistochemistry) from the same section of a human tumor resection (im-munostained simultaneously). HSF1-expressing cells stain brown, and HSF1-negative cells stain blue from the toluidine blue counterstain. Scale bar, 50 mm. (D)HSPA1A mRNA transcript levels are elevated in colorectal adenocarcinomas with high-grade aneuploid karyotypes. Data from three MIN and nine CIN coloncancer cell lines from the GSK Cancer Cell Line Genomic Profiling Data as described in themethods. Box plot, bar is median, and whiskers are min andmax (threecell lines in MIN category in triplicate 9 cell lines in CIN category in triplicate). (E) RT-PCR analysis ofHSPA1AmRNA levels in the indicated euploid nontransformedcell lines, and MIN and CIN cancer cell lines. Error bars indicate mean T SD of triplicates. (F) Effect of RHT on the proliferation of a panel of cell lines with high-gradeaneuploid karyotypes (CIN lines: Caco2, HT29, SW403, SW480, and SW620); near-euploid karyotypes with microsatellite instability (MIN lines: HCT-116, HCT-15,DLD-1, SW48, and LoVo); or nontransformed colon epithelial cell lines with a euploid chromosomal number (CCD112CoN and CCD841CoN), (mean T SD, n = 3replicates, ***P < 0.001, two-way ANOVA) treated for 72 hours.

www.sciencemag.org SCIENCE VOL 341 19 JULY 2013 1238303-7

RESEARCH ARTICLE

HSF1-bound genes among the genes differen-tially expressed after treatment of MCF7 cells withtranslation elongation inhibitors was conductedby using GSEAv2.08 software (42). HSF1-boundgenes in MCF7 cells were defined as those genesbound in at least two of the four data sets [twodata sets from this study and two from (8)].

Evaluation ofHSPA1AmRNA levels was per-formed by using data from the GlaxoSmithKline(GSK) Cancer Cell Line Genomic Profiling Datahttps://cabig.nci.nih.gov/community/caArray_GSK-data.MIN lines usedwereHCT15,LS174T, SW48.CIN lines usedwereNCIH508,NCIH747, SW1116,SW1417, SW403, SW480, SW620,T84, andSW948.

ChiP-Seq and ChIP-PCRDescribed in supplementary materials, materialsand methods.

ImmunoblotDescribed in supplementary materials, materialsand methods.

Fig. 7. Rocaglates suppress tumor growth, HSPA8 mRNA levels andglucose uptake in vivo. (A) Scatter plot of IC50 values of the growth of a diversepanel of cell lines treated with RHT. Cells were treated for 5 days. Red indicateshematopoietic cancer lines, dark blue indicates solid tumor cell lines, and lightblue indicates euploid nontransformed cells. (B) mRNA levels of HSPA8, TXNIP,and control housekeeping genes in M0-91 cells treated with RHT. Mean of du-plicates are reported. (C) Glucose uptake of of IR Dye 800CW2-deoyglucose (2-DG)in M0-91 cell lines treated with RHT. Imaging was performed by using LICOR.(Right) Quantitation of measured intensity (error bars indicate mean T SEM; P <0.005, two-tailed t test, n = 4 replicates). (D) Plot of the tumor volume of M0-91acute myeloid leukemia xenografts treated with vehicle or RHT. The mean tumorvolume (in cubic millimeters) is plotted over time. Mice were treated with sub-cutaneous injections starting on day 18 after implantation [either vehicle alone or

RHT (1mg/kg), on days marked by downward pointing arrows]. Eight mice were ineach treatment group (error bars indicate mean T SEM; P < 0.0001). (E) RT-PCRanalysis of HSPA8 and TXNIP mRNA levels from tumor xenografts after a singletreatment of either vehicle or RHT (1mg/kg, subcutaneous; fivemice in each group).Tumors were harvested 4 hours after treatment (error bars indicate mean T SEM;HSPA8, P < 0.005; TXNIP, P < 0.0005, two-tailed t test, n = 5 replicates). (F)Representative image of epifluorescence of IRDye 800CW 2-deoxyglucose (2-DG)uptake in M0-91 xenografts. Mice bearing tumors were treated with vehicle orRHT (1 mg/kg) as described in the supplementary materials, materials andmethods; four mice in each group. Images were acquired 48 hours after the lasttreatment. (Right) Quantitation of measured radiant efficiency from epifluo-rescence of IRDye 800CW 2-DG from images of M0-91 xenografts (error barsindicate mean T SEM; P = 0.031, two-tailed t test, n = 4 replicates).

19 JULY 2013 VOL 341 SCIENCE www.sciencemag.org1238303-8

RESEARCH ARTICLE

LINCS AnalysisTo identify chemical and genetic modulators thatare correlated with HSF1 inactivation, we queriedLINCS supported by the NIH Common Fund.This resource at the Broad Institute is a massiveexpression-profiling initiative to catalog the cel-lular consequences of both small-molecule andgenetic perturbations. The expression data weregenerated by using a high-throughput Luminex-based assay (Luminex, Austin, TX) as describedpreviously (43). Whole-genome expression pro-files are inferred from changes in the expressionof 1000 landmark genes. The changes in geneexpression resulting from each of the geneticand chemical perturbations are rank-ordered fromhighest to lowest according to their differentialexpression relative to control treatments. Changesin gene expression caused by a novel perturbation(the “query” signature) can then be comparedwiththe cataloged expression profiles. Profiles thatare positively correlated with the query signatureare given a positive score, whereas profiles that arenegatively correlated to the query signature aregiven a negative score.

For the analysis, we generated an HSF1 in-activation signature (table S4) of the 50 genesmost positively regulated (reduced expression uponHSF1 depletion with shRNA) and 10 genes mostnegatively regulated (increased expression uponHSF1 depletionwith shRNA) in the breast cancercell lines,MCF7 andBPLER (44) [average of thedifference between the ha6 shRNA and scram-bled shRNA control values between the two celllines (8)], that were also bound by HSF1 in ourChIP-seq experiments. This signature was usedto query all 161,636 shRNA and compound sig-natures (collapsed from a total of 614,216 indi-vidual profiles from at least three biologicalreplicates) in the LINCS data set produced innine cell lines (MCF7, breast cancer; HT29, co-lon cancer; HEPG2, hepatoblastoma; A549, lungcancer; HCC515, lung cancer; A375, melanoma;PC3, prostate cancer;VCAP, prostate cancer;HA1E,immortalized but nontransformed kidney epithe-lium). A connectivity score was assigned to eachof the expression profiles from the 161,636 per-turbations on the basis of a weighted Kolmogorov-Smirnov statistic as previously described (42, 43).GSEA (42) was performed on this rank-orderedlist in order to determine gene or chemical classesthat weremost enriched among the positively andnegatively connected signatures. The sets analyzedby means of GSEA encompassed the shRNAscorresponding to the genes comprising all 186KEGG pathway gene sets. The sets also included110 chemical classes grouped according to theAnatomical Therapeutic Chemical (ATC) Classi-fication System. In addition, we added a set com-posed of elongation initiation factors. Statisticalsignificance was tested by using 100 random setssize-matched to the set being tested.

Reporter cell linesY9 reporter NIH3T3 cells (45) were infectedwith lentivirus for the doxycycline regulatable

pTRIPZ-nonsilencing construct (RHS4743).These cells were heat shocked and incubated withdoxycycline and then sorted by use of flow cy-tometry to isolate strong enhanced GFP (eGFP)and Turbo RFP (tRFP) expressors. Sorting wasrepeated twice for enrichment. Noninduced cellswere sorted to remove cells expressing eGFP andtRFP at baseline to make R4.1.B4 cells.

Tomake the high-throughput screening cell lineNIH3T3HGL, the parent vector LV-eGFPfLUCwas modified by removing the cytomegaloviruspromoter and introducing a 470-bp fragment ofthe human HSP70B′ construct upstream of theeGFP promoter. NIH3T3 cells were infected withlentivirus generated from this construct, and thehigh eGFP expressors were isolated by means offlow cytometry after heat shock.

High-Throughput Small-Molecule ScreenDescribed in supplementary materials, materialsand methods.

Dual Reporter Cell AssayDescribed in supplementary materials, materialsand methods.

Rocaglamide/Rocaglate DerivativesRocaglamide/rocaglate derivativeswere preparedwith total synthesis methods as previously de-scribed (12, 13).

Nanostring/nCounter AnalysisThe cells were lysed at a concentration of 10,000cells/mLwith RTL buffer [Qiagen (Valencia, CA)catalog no. 79216] and dissociated using a cellshredder (Qiagen catalog no. 79656). The totalRNA in 5 ml of lysate was hybridized with thecapture and reporter probes overnight at 65°Cand processed according to the nCounter recom-mended protocol. Target/probe complexes wereimmobilized in nCounter Cartridges for data col-lection using an nCounter Digital. The data wasanalyzed according to the manufacture’s guide-lines. All data were normalized to the expres-sion levels of house keeping genes (ACTB, B2M,GAPDH, GUSB, HPRT1, and RPL10).

S35 LabelingMCF7 cells were grown to confluence in six-welldishes in standard Dulbecco’s minimum essentialmedium (DMEM) (+10% fetal bovine serum).The cells were rinsed twice in 1× phosphate-buffered saline (PBS) and then placed in DMEMwithout methionine or cysteine [Life Technolo-gies (Guilford, CT) no. 21013024] for 30 min.EasyTagS35protein labelingmix (NEG772002MC)was added for 15 min. The cells were rinsedtwice in 1X PBS and then lysed in TNEK buffer.Cell lysateswere prepared in TNEKbuffer {50mMTris, pH7.4;NP-401%;EDTA2mM;KCl 200mMand protease inhibitor cocktail [Roche Diagnos-tics (Indianapolis, IN) catalog no. 11836153001]}.Samples (15 mg total protein/lane) were analyzedby means of SDS–polyacrylamide gel electro-phoresis. The gelwas incubated for 10min in 0.7M

sodium salicylate and 10% glycerol. The driedgel was used to expose film. Counts per minutewere assessed by using a scintillation counter.

Glucose Uptake500,000 cells plated in 24-well tissue cultureplates were treated for 8 hours with RHT. Thecells were then washed in 1× PBS (×2) andplaced in glucose-free and serum-free DMEMfor 20 min. Glucose uptake was measured byusing 3H-2-Deoxyglucose (3H-2DG), incubationfor 15 min (final 1 mCi/ml, 50 mM 2DG). Thecells were washedwith 1ml of cold 1× PBS (×2),lysed with 0.2 M sodium hydroxide, and thencounted by use of scintillation. Experiments wereperformed in triplicate. Parallel treated culturescells were stained with Sytox-green (Invitrogen,catalog no. S7020) for normalization. Each an-alysis was performed three times. The SEM isdisplayed.

Glucose uptake was also measured by usingIRD800 2-deoxyglucose (Fig. 7C). M0-91 cellswere washed and resuspended in glucose-freemedium (106 cells/ml). Cells were dispensed(100 mL/well) in triplicate wells and incubatedfor 20minwith 5 mMof the IRDye 800CW2-DGOptical Probe [LI-COR Biosciences (Lincoln, NB)catalog no. 926-08946]. The cellswere thenwashed×4 with cold PBS, and the signal was acquiredby using an infrared Odyssey imaging system(LI-COR Biosciences).

Lactate Production500,000 cells were plated in 24-well tissue cul-ture plates and were treated for 8 hours. withRHT. After two washes with 1× PBS, the cellswere incubated for 30 min at 37°C in 500 ml offilter sterilized 1× Krebs buffer (126 mMNaCl,2.5mMKCl, 1.2mMNaH2PO4, 1.2mMMgCl2,2.5mMCaCl2, 10mMGlucose, 25mMNaHCO3,10 mM HEPES-KOH pH 7.4). The supernatantwas collected, and the lactate was measured witha Lactate Assay Kit [BioVision (Milpitas, CA) cat-alog no. K-607) according to the manufacturer’sguidelines. Parallel treated cultures cellswere stainedwith SytoxGreen (Invitrogen, catalog no. S7020)for normalization. Each analysis was performedthree times. The SEM is displayed.

Cell Viability AssayRelative cell growth and survival were measuredin 96-well microplate format by using the fluo-rescent detection of resazurin dye reduction as anendpoint (544 nm excitation and 590 nm emis-sion). Two thousand adherent cells and 10,000suspension cells were plated 24 hours before com-pound exposure (for 72 hours). Each analysiswas performed three times. For all bar graphs, theSEM is displayed, unless indicated otherwise.

ImmunohistochemistryParaffin blocks of human colon adenocarcinomatissue were obtained from the archives of BWHin accordance with the regulations for excess tis-sue use stipulated by theBWH institutional review

www.sciencemag.org SCIENCE VOL 341 19 JULY 2013 1238303-9

RESEARCH ARTICLE

board. Immunohistochemistry for HSF1was per-formed as previously described (8).

Drug Metabolism and Pharmacokinetic StudiesDescribed in supplementary materials, materialsand methods.

Xenograft Experiment5e7 M0-91 cells were implanted with Matrigel(BD Biosciences, San Jose, CA) subcutaneously inthe right inguinal region of NOD-SCID mice.When themean tumor volume reached 100mm3,RHT formulated in hydroxypropyl beta-cyclodextrinwas administered by means of subcutaneous par-enteral administration (1 mg/kg) according tothe treatment schedule shown in Fig. 7D. Tumorsize was measured twice each week by a labmember (M.D.) whowas blinded to the treatmentgroups. There were eight mice in each treatmentgroup (RHT-treated or vehicle-treated). Bodyweightwas measured twice each week. Body conditionscore (BCS) as monitored by facility vets did notgo below 2+ in either treatment group throughoutthe experiment.

In Vivo Glucose Uptake ExperimentM0-91 cells were inoculated into the inguinal re-gion of NOD-SCID mice. Seventeen days later,themicewere treatedwith a dose of RHT (1mg/kg;4 mice) or vehicle control (4 mice). Four hourslater, the mice were given retro-orbital injectionsof 100 ml IRDye 800CW 2-DG Optical Probe(10 nmol; LI-CORBiosciences, no. 926-08946),and then an additional 4 hours later, these micewere again treated with RHT (1mg/kg) or vehiclecontrol. Thirty-six hours after the last RHT dose,mice were imaged (IVIS; excitation 745 nm, emis-sion 800 nm). Data were analyzed by using LivingImage software (PerkinElmer, Waltham, MA).

Real-Time PCRRNAwas purified with RNEasy columns (Qiagen,catalog no. 74104). Quantitative PCR to evaluatemRNA levels was performed by using RT2 SYBRGreenqPCRMastermix (SABiosciences) andprimerassay pairs (SABiosciences, Valencia, CA) on a7900HTABI Detection System.

References and Notes1. D. Ruggero, P. P. Pandolfi, Nat. Rev. Cancer 3, 179–192

(2003). doi: 10.1038/nrc1015; pmid: 126126532. J. C. Chan et al., Sci. Signal. 4, ra56 (2011).

doi: 10.1126/scisignal.2001754; pmid: 218786793. R. J. White, Nat. Rev. Mol. Cell Biol. 6, 69–78 (2005).

doi: 10.1038/nrm1551; pmid: 156880684. K. M. Hannan et al., Mol. Cell. Biol. 23, 8862–8877

(2003). doi: 10.1128/MCB.23.23.8862-8877.2003;pmid: 14612424

5. D. Silvera, S. C. Formenti, R. J. Schneider, Nat. Rev. Cancer10, 254–266 (2010). doi: 10.1038/nrc2824;pmid: 20332778

6. P. P. Roux, I. Topisirovic, Cold Spring Harb. Perspect. Biol. 4,a012252 (2012). doi: 10.1101/cshperspect.a012252;pmid: 22888049

7. F. U. Hartl, M. Hayer-Hartl, Science 295, 1852–1858(2002). doi: 10.1126/science.1068408; pmid: 11884745

8. M. L. Mendillo et al., Cell 150, 549–562 (2012).doi: 10.1016/j.cell.2012.06.031; pmid: 22863008

9. L. Whitesell, S. Lindquist, Expert Opin. Ther. Targets 13,469–478 (2009). doi: 10.1517/14728220902832697;pmid: 19335068

10. M. E. Bordeleau et al., J. Clin. Invest. 118, 2651–2660(2008). pmid: 18551192

11. J. M. Chambers et al., Org. Lett. 15, 1406–1409(2013). doi: 10.1021/ol400401d; pmid: 23461621

12. C. M. Rodrigo, R. Cencic, S. P. Roche, J. Pelletier,J. A. Porco Jr., J. Med. Chem. 55, 558–562 (2012).doi: 10.1021/jm201263k; pmid: 22128783

13. S. P. Roche, R. Cencic, J. Pelletier, J. A. Porco Jr.,Angew. Chem. Int. Ed. Engl. 49, 6533–6538 (2010).doi: 10.1002/anie.201003212; pmid: 20687060

14. X. Jin, D. Moskophidis, N. F. Mivechi, Cell Metab. 14,91–103 (2011). doi: 10.1016/j.cmet.2011.03.025;pmid: 21723507

15. C. Dai, L. Whitesell, A. B. Rogers, S. Lindquist, Cell 130,1005–1018 (2007). doi: 10.1016/j.cell.2007.07.020;pmid: 17889646

16. H. Parikh et al., PLoS Med. 4, e158 (2007). doi: 10.1371/journal.pmed.0040158; pmid: 17472435

17. C. A. Stoltzman et al., Proc. Natl. Acad. Sci. U.S.A. 105,6912–6917 (2008). doi: 10.1073/pnas.0712199105;pmid: 18458340

18. S. Y. Kim, H. W. Suh, J. W. Chung, S. R. Yoon, I. Choi,Cell. Mol. Immunol. 4, 345–351 (2007). pmid: 17976314

19. C. Dai et al., J. Clin. Invest. 122, 3742–3754 (2012).doi: 10.1172/JCI62727; pmid: 22945628

20. Y. C. Tang, B. R. Williams, J. J. Siegel, A. Amon, Cell 144,499–512 (2011). doi: 10.1016/j.cell.2011.01.017;pmid: 21315436

21. B. R. Williams et al., Science 322, 703–709 (2008).doi: 10.1126/science.1160058; pmid: 18974345

22. S. Santagata et al., Proc. Natl. Acad. Sci. U.S.A. 108,18378–18383 (2011). doi: 10.1073/pnas.1115031108;pmid: 22042860

23. F. Fang, R. Chang, L. Yang, Cancer 118, 1782–1794(2012). doi: 10.1002/cncr.26482; pmid: 22009757

24. L. Alinari et al., Clin. Cancer Res. 18, 4600–4611 (2012).doi: 10.1158/1078-0432.CCR-12-0839; pmid: 22791882

25. R. Cencic et al., PLoS ONE 4, e5223 (2009).doi: 10.1371/journal.pone.0005223; pmid: 19401772

26. D. M. Lucas et al., Blood 113, 4656–4666 (2009).doi: 10.1182/blood-2008-09-175430; pmid: 19190247

27. M. Okabe et al., Leuk. Res. 19, 933–943 (1995).doi: 10.1016/0145-2126(95)00039-9; pmid: 8632663

28. L. Meng, V. L. Gabai, M. Y. Sherman, Oncogene 29,5204–5213 (2010). doi: 10.1038/onc.2010.277;pmid: 20622894

29. S. Santagata et al., ACS Chem. Biol. 7, 340–349 (2012).doi: 10.1021/cb200353m; pmid: 22050377

30. K. L. Scott et al., Cancer Cell 20, 92–103 (2011).doi: 10.1016/j.ccr.2011.05.025; pmid: 21741599

31. V. L. Gabai et al., Mol. Cell. Biol. 32, 929–940 (2012).doi: 10.1128/MCB.05921-11; pmid: 22215620

32. S. D. Westerheide, J. Anckar, S. M. Stevens Jr.,L. Sistonen, R. I. Morimoto, Science 323, 1063–1066(2009). doi: 10.1126/science.1165946; pmid: 19229036

33. O. Brandman et al., Cell 151, 1042–1054 (2012).doi: 10.1016/j.cell.2012.10.044; pmid: 23178123

34. J. S. Hahn, Z. Hu, D. J. Thiele, V. R. Iyer, Mol. Cell. Biol.24, 5249–5256 (2004). doi: 10.1128/MCB.24.12.5249-5256.2004; pmid: 15169889

35. A. L. Hsu, C. T. Murphy, C. Kenyon, Science 300, 1142–1145(2003). doi: 10.1126/science.1083701; pmid: 12750521

36. S. D. Chou, T. Prince, J. Gong, S. K. Calderwood,PLoS ONE 7, e39679 (2012). doi: 10.1371/journal.pone.0039679; pmid: 22768106

37. S. M. Hensen et al., Cell Stress Chaperones 17, 743–755(2012). doi: 10.1007/s12192-012-0347-1; pmid: 22797943

38. T. Peng, T. R. Golub, D. M. Sabatini, Mol. Cell. Biol. 22,5575–5584 (2002). doi: 10.1128/MCB.22.15.5575-5584.2002; pmid: 12101249

39. L. Whitesell, S. Santagata, N. U. Lin, Curr. Mol. Med. 12,1108–1124 (2012). doi: 10.2174/156652412803306657;pmid: 22804235

40. N. L. Solimini, J. Luo, S. J. Elledge, Cell 130, 986–988(2007). doi: 10.1016/j.cell.2007.09.007; pmid: 17889643

41. J. Lamb et al., Science 313, 1929–1935 (2006).doi: 10.1126/science.1132939; pmid: 17008526

42. A. Subramanian et al., Proc. Natl. Acad. Sci. U.S.A. 102,15545–15550 (2005). doi: 10.1073/pnas.0506580102;pmid: 16199517

43. D. Peck et al., Genome Biol. 7, R61 (2006).doi: 10.1186/gb-2006-7-7-r61; pmid: 16859521

44. T. A. Ince et al., Cancer Cell 12, 160–170 (2007).doi: 10.1016/j.ccr.2007.06.013; pmid: 17692807

45. T. J. Turbyville et al., J. Nat. Prod. 69, 178–184(2006). doi: 10.1021/np058095b; pmid: 16499313

Acknowledgements: We thank T. Volkert, J. Love, S. Gupta,and the WIBR–Genome Technology Core for sequencingsupport; S. Malstrom (Koch Institute for Integrative CancerResearch) for assistance with in vivo imaging; G. Bell, P. Thiru,and A. Lancaster for assistance with informatics analysis; theConnectivity Map team at the Broad Institute for generationof the LINCS data set and query tools; Joe Negri and theMLPCN team at the Broad Institute for chemical screening;and M. Duquette for assistance with animal experiments.We also thank C. Rodrigo (Boston University) for compoundsynthesis. We thank the Lindquist laboratory for helpfuldiscussions and suggestions. The work was supported bythe Johnson & Johnson’s Corporate Office of Science andTechnology focused funding program (L.W.), the MarbleFund (S.L.), and NIH R01 CA175744-01 (L.W.). The MLPCNscreen was supported by R03 MH086465-01 and R03DA027713-01 to L.W. This work was supported by the NIHCommon Fund’s LINCS program (5U54HG006093, “Largescale gene expression analysis of cellular states”) to T.R.G.J.A.P. Jr. is supported by R01 GM073855. S.L. is anInvestigator of the Howard Hughes Medical Institute. M.L.M.was supported by American Cancer Society New EnglandDivision–SpinOdyssey (PF-09-253-01-DMC). S.S. is supportedby NIH (K08NS064168), the Brain Science Foundation, theAmerican Brain Tumor Association, the Beez Foundation,the V Foundation, and the Jared Branfman Sunflowers forLife Fund. The expression profiling and ChIP-Seq data aredeposited in GEO (GSE45853). The MLPCN chemical screeningdata are deposited in Pubchem (AID: 2118).

Supplementary Materialswww.sciencemag.org/content/full/341/6143/1238303/suppl/DC1Materials and MethodsFigs. S1 to S9Reference (46)Tables S1 to S5

25 March 2013; accepted 24 May 201310.1126/science.1238303

19 JULY 2013 VOL 341 SCIENCE www.sciencemag.org1238303-10

RESEARCH ARTICLE